[Paper Review] BEVPlace: Learning LiDAR-based Place Recognition using Bird's Eye View Images

Luo Lun, Shuhang Zheng+5 more|arXiv (Cornell University)|Feb 28, 2023

Robotics and Sensor-Based Localization4 citations

TL;DR

This paper proposes BEVPlace, a LiDAR-based place recognition method that uses bird's eye view (BEV) images to improve robustness to viewpoint changes and scene variations. By leveraging group convolution for rotation-equivariant feature extraction and NetVLAD for global feature aggregation, BEVPlace achieves state-of-the-art recall rates (95.3% mean AR@1 on benchmark datasets), strong generalization, and accurate 6-DoF position estimation via learned feature-to-geometry distance correlation.

ABSTRACT

Place recognition is a key module for long-term SLAM systems. Current LiDAR-based place recognition methods usually use representations of point clouds such as unordered points or range images. These methods achieve high recall rates of retrieval, but their performance may degrade in the case of view variation or scene changes. In this work, we explore the potential of a different representation in place recognition, i.e. bird's eye view (BEV) images. We observe that the structural contents of BEV images are less influenced by rotations and translations of point clouds. We validate that, without any delicate design, a simple VGGNet trained on BEV images achieves comparable performance with the state-of-the-art place recognition methods in scenes of slight viewpoint changes. For more robust place recognition, we design a rotation-invariant network called BEVPlace. We use group convolution to extract rotation-equivariant local features from the images and NetVLAD for global feature aggregation. In addition, we observe that the distance between BEV features is correlated with the geometry distance of point clouds. Based on the observation, we develop a method to estimate the position of the query cloud, extending the usage of place recognition. The experiments conducted on large-scale public datasets show that our method 1) achieves state-of-the-art performance in terms of recall rates, 2) is robust to view changes, 3) shows strong generalization ability, and 4) can estimate the positions of query point clouds. Source codes are publicly available at https://github.com/zjuluolun/BEVPlace.

Motivation & Objective

• To address the limitations of existing LiDAR-based place recognition methods that degrade under viewpoint changes and scene variations.
• To explore the potential of bird's eye view (BEV) images as a more robust representation for LiDAR point clouds.
• To develop a rotation-invariant network that maintains high performance under viewpoint variations and scene changes.
• To extend place recognition beyond retrieval by enabling accurate 6-DoF position estimation using learned feature-to-geometry distance correlation.

Proposed method

• The method projects LiDAR point clouds into BEV images, which are more stable under sensor translations and rotations compared to range images or unordered point clouds.
• A group convolution-based network is designed to extract rotation-equivariant local features from BEV images, enhancing robustness to viewpoint changes.
• NetVLAD is applied to aggregate local features into a global, rotation-invariant descriptor for place retrieval.
• A learned correlation model maps the L2 distance between BEV features to the geometric distance between point clouds, enabling position estimation.
• The framework is trained end-to-end on public datasets like KITTI and Oxford RobotCar, with inference achieving real-time performance (~30ms per query).

Experimental results

Research questions

• RQ1 Can BEV images serve as a more robust representation than range images or unordered point clouds for LiDAR-based place recognition under viewpoint variations?
• RQ2 How can a deep learning model achieve rotation invariance in BEV image-based feature learning for improved generalization?
• RQ3 Can the distance between BEV features be reliably correlated with the geometric distance between point clouds to enable position estimation?
• RQ4 Does the proposed BEVPlace method outperform state-of-the-art methods in terms of recall, robustness, and generalization across diverse datasets?

Key findings

• BEVPlace achieves a mean Top-1 recall of 95.3% on the benchmark dataset, outperforming state-of-the-art methods including transformer-based models like MinkLoc3Dv2 and ResNet34+NetVLAD.
• The method demonstrates strong generalization, maintaining high performance on the ALITA dataset when trained on KITTI, while other methods show significant performance drops.
• BEVPlace achieves 96.5% AR@1 on the Oxford dataset and 96.9% on the U.S. dataset, showing superior robustness under viewpoint changes and scene variations.
• The feature-to-geometry distance correlation enables accurate position estimation, with cumulative translation error distribution showing superior performance under rotation perturbations.
• Runtime evaluation shows that place recognition takes ~30ms per query on an RTX 3090 GPU, enabling real-time localization at 10Hz, suitable for SLAM systems.